Wove

Local, autonomous developer agent with built-in browser and computer vision.

Wove is not a terminal with an AI chat. It's an autonomous agent that sees your screen, browses the web, runs commands, delegates sub-tasks, and orchestrates multi-step workflows — all running locally on your machine.

Built on the Wave Terminal engine (Go + Electron, Apache 2.0).

Download

macOS (Apple Silicon / M1+):
Wove-darwin-arm64-0.2.0.dmg ·
zip

macOS (Intel):
Wove-darwin-x64-0.2.0.dmg ·
zip

Linux and Windows builds coming soon.

All releases

What makes Wove different

Core Capabilities

Computer Vision & Web Automation

Wove sees web pages through CDP-based snapshots with Set-of-Mark numbered markers. The agent identifies interactive elements, reads content, and automates the browser:

• web_capture — visual snapshot with numbered element markers and CSS selectors
• web_click / web_mouse_click — click by selector or coordinates (CDP-based, works in iframes)
• web_type_input — type into inputs with framework event dispatch
• web_exec_js — execute JavaScript in page context
• web_open / web_navigate — open and navigate browser widgets
• web_read_text / web_read_html — extract content by CSS selector
• web_seo_audit — full page audit (JSON-LD, OG, meta, headings, links)

Sub-task Orchestration

Complex tasks (audits, migrations, multi-file refactors) are split into isolated sub-tasks. Each sub-task runs in a new tab with fresh context and full tool access, preventing context window overflow.

Parent task: "Full SEO audit of example.com"
    |
    +-- Sub-task 1: Technical audit → /tmp/seo/01-technical.md
    +-- Sub-task 2: Content quality → /tmp/seo/02-content.md
    +-- Sub-task 3: Schema markup  → /tmp/seo/03-schema.md
    |
    v
Parent reads output files → generates consolidated report

MCP Integration

Auto-detects .mcp.json in your project. AI gets direct access to your database schema, documentation, logs — any MCP-compatible data source.

AI Planning System

Multi-step execution plans with:

• Concrete file paths and pattern references
• Auto-appended steps: lint, review, test
• Live progress panel
• Plans survive restarts

Project Intelligence

Reads WAVE.md, CLAUDE.md, .cursorrules automatically:

• Tech stack injected into every request
• Critical rules always present
• Project tree on first message
• Smart filtering by technology

Skills System

Extensible skill definitions (SKILL.md files) with:

• invoke_skill tool for on-demand loading
• Slash command autocomplete in AI input (/seo-audit, /seo-technical, etc.)
• Skills run autonomously end-to-end without confirmation prompts

Code Quality Guardrails

Wove enforces multiple layers of code quality to ensure generated code matches your project's conventions:

• Read-before-edit — the AI must read a file before modifying it (enforced at tool level, not just prompt)
• Architecture matching — find 2-3 similar files, study their patterns, match the style exactly
• Code discipline — no unnecessary refactoring, no premature abstractions, no gold-plating beyond what was asked
• Security awareness — OWASP-aware: parameterized queries, output escaping, no secrets in code
• Self-review — after every write, re-read the file and compare with the reference for style drift
• Verification — run linters and tests after code changes; never claim "done" without evidence
• Warm context — technology-specific conventions from WAVE.md are re-injected when editing files of that type
• Sibling references — when creating new files, an existing file from the same directory is automatically attached as a style reference
• Project instructions override — WAVE.md/CLAUDE.md rules explicitly override the AI's default behavior

Smart Context Management

Long coding sessions stay fast — Wove automatically manages the AI's context window:

• Screenshot stripping — base64 images shown to the model once, then replaced with text element list in history
• Tool result compaction — old tool outputs are truncated in API requests while chatstore keeps full data for UI
• Conversation compaction — automatic cleanup when context grows too large
• Compressed screenshots — 512px JPEG at quality 30 (~15-25KB instead of 50-80KB)
• Model switch protection — switching models auto-clears chat to prevent format mismatches

Widget Ownership & Cleanup

Agent tracks which terminals and browsers it created. Closes them when done. Cannot close user's pre-existing widgets.

Multi-Model Support (BYOK)

Session History

Chat history persists per tab across restarts, with a visual banner showing previous work.

Getting Started

Requirements

• macOS 11+, Windows 10 1809+, or Linux (glibc-2.28+)
• Node.js 22 LTS
• Go 1.25+
• Task build tool

Build and run

git clone https://github.com/mits-pl/wove.git
cd wove
task init
task dev

Set up an AI model (required)

Wove needs an AI model to work. The fastest way to get started:

1. Open Wove and find the AI panel on the right side (or press Alt+Shift+A)
2. Right-click anywhere in the AI panel to open the context menu
3. Go to Quick Add Model and pick your provider
4. Paste your API key when prompted — done

API keys are stored securely in an encrypted local keystore — never in config files or plain text.

Recommended for coding: Claude Sonnet 4.6 or GPT-5.1 with tool support enabled.

You can also click Configure Modes in the same context menu to open the JSON editor and add or modify models directly.

Advanced: manual model configuration

For custom endpoints or models not in Quick Add, create ~/.config/woveterm/waveai.json:

{
  "my-model": {
    "display:name": "My Model",
    "ai:apitype": "anthropic-messages",
    "ai:model": "claude-sonnet-4-6",
    "ai:endpoint": "https://api.anthropic.com/v1/messages",
    "ai:apitokensecretname": "my_api_key",
    "ai:capabilities": ["tools", "images", "pdfs"]
  }
}

Supported ai:apitype values: anthropic-messages, openai-responses, openai-chat, google-gemini

Project Instructions

Create WAVE.md in your project root:

## Project
My App — Laravel 11, Inertia.js, Vue 3

## Conventions
- Always use Form Request classes for validation
- Use Eloquent scopes, not raw queries
- Run vendor/bin/pint after changes

Architecture

User message
    |
    v
[Project Stack] -> "Laravel + Inertia + Vue"
[Critical Rules] -> "must write tests, must use PHPDoc"
[MCP Context]   -> database schema, app info
[Active Plan]   -> what step to execute next
    |
    v
Agent: plan -> read code -> write -> review -> test -> lint
    |
    +-- needs web data? -> web_open -> web_capture -> web_read_text
    +-- complex task? -> run_sub_task (isolated tab)
    +-- done? -> close_widget (cleanup)

Built On

Wove is built on Wave Terminal by Command Line Inc., licensed under Apache License 2.0.

See MODIFICATIONS.md for a complete list of changes from upstream.

Contributing

Issues and PRs welcome. See CONTRIBUTING.md.

Support

If you find Wove useful, consider buying us a coffee:

License

Apache License 2.0. See LICENSE.